๐๐ DaSiWa-WAN 2.2 I2V 14B SynthSeduction v9 | Lightspeed | GGUF ๐๐
GGUF Version of the model.
This is a WAN 2.2 Model: You will need one pair of High + Low.
Version overview: https://civarchive.com/articles/23495/dasiwa-model-versions-and-timeline
๐ฎ Key Features:
๐ฅ LoRA-Free Generations
Generate high-quality videos without stacking Wan 2.2 LoRAs (unless you want adding spacial styles/concepts).โ๏ธFast: 4 step generation
๐ซQuality motions (less slowdowns, no pixelated hyper-motion)
๐ NSFW and SFW + Extreme versatile (more build in concepts):
Enhanced anatomy + poses + framing
Better understanding of sexual concepts
๐ช Better prompt responsiveness
๐ Better understanding of anime/manga style composition
๐ชก Q8 (FP16 base) precision
๐ซ Do not use any extra speed-up (low step) LoRAs, this is baked in already
โ Optimisations
๐ CFGZeroStar patch (better results and prompt adherence)
๐ฐ Baked Latest Distillation (r64-1022)
โฉ SkyReels V2 for long consistent videos
๐Additional concept optimisation
๐ฌ Reward Attention (more realistic movements)
๐ฎ VTA (Better completions, transitions, input understanding)
โจ No or extreme low transformation of details with anime/realistic images (lips, eyes, ears, breasts, genitals, ...)
๐ Faster and refined motions - As close to RL motion-speeds as I could get with speed-up tech without morphing
๐ Guesstimate better details
๐Even less tries for good results
๐งฉ Raised compatibility with LoRAs
๐ผ๏ธ Usable with your preferred/custom CLIP (if compatible)
๐ตโ๐ซ Reduced hallucinations
๐ Zero prompt results capable
๐ซ Excluded CLIP
๐Workflow
Make sure to checkout my easy to use Workflows!
๐LoRA's
Try first without additional LoRAs!
But: This checkpoint is not meant to replace all LoRAs, it is meant to:
Perform better overall at his own
As easy as possible to use
With LoRAs to be absolutely awesome
๐ชงAnnouncement
โ ๏ธ Read the corresponding announcement.
๐ข Make sure to check it out for in-depth information and a complex comparison!
๐ New to WAN 2.2 I2V? - Check out my guide.
๐ ๏ธ Recommended Settings
Steps: 4
CFG: 1
Sampler/Scheduler: Euler/Simple, Euler/linear_quadratic
Resolution up to 720p (native quality).
Add other LoRAs with 0.3 - 0.6 at first
16 or 24 fps, 81 or 97 frames ~ 5s
Dependencies
๐ซ Speed + Examples
Q8 checkpoint - On 16GB VRAM, 64GB RAM, 4 steps, cfg 1, 81 frames
368p: 120 sec
480p: 160 sec
576p: 220 sec
608p: 340 sec
672p: 680 sec
720p: 730 sec
Most examples are without any additional LoRAs
With LoRAs are for testing the compatibility
Initial anime-like example images made by me are also made with my model ๐ก๏ธ๐ DaSiWa-Illustrious-XL ๐๐ก๏ธ
Other models for realistic reasoning
๐ฉป Known issues
๐ซฆ This Synth is adorable!
Tell me ๐ซต๐ซข
Approximate expected quality from quantization
Refer to this comparison article.
This are my tests compared to a full fp16 safetensor checkpoint taking prompt- and visual satisfaction into account on my DaSiWa checkpoints.
โ ๏ธ Do not compare this with the unofficial quants of my checkpoint made by others, they are based on FP8 and not FP16 like my quants!
๐ฉบ Fixes & Feedback
If you use LoRAs, try to respect the LoRA training triggers and try some versatile descriptions, most LoRAs will work with 0.3-0.6 (start with 0.3)
Raise LoRAs in little steps +0.1
Do not mass add LoRAs, just add 1 or 2 (x2 High+Low)
Negative prompting do not work with cfg 1, that's a limitation of speed-ups with cfg 1 (except you use NAG)
Low resolution (e.g.384x576) are for fast samples and will blur fine details, do a higher resolution if you want clear details
๐ชงโ Test your comfyui-backend with this absolute basic test-workflow before asking about errors.
๐ค Why I Made This
I was tired of using all these massive list of LoRAs, just to get a remotely good result after 10 generations, consuming hours of time. Also was not satisfied with the results other checkpoints can achieve across the board.
So I can just make my videos with 1 or 2 concept LoRAs without pushing 6 till 10 LoRAs (Low/High) into a generation.
This checkpoint is also my personal playground.
Closing words
๐คฉ I want to thank all the fantastic other creators who made super nice LoRAs and concepts to play with! Support that awesome creators by using their LoRAs and post to their gallery and share the meta-data!
โ ๏ธ I made all this with permissions or open-source resources (the time it is incorporated).
I share as much insights as I can without compromising my work. I'm doing this for fun as my hobby and just do not want my hobby to be destroyed.
More details can be obtained in the corresponding announcements!
If you would like to contribute in my awesome (๐) checkpoint or willing to share resources I'll gladly give credit! Just contact me!
โ All credits / resources are mentioned inside the announcements! - Since different versions may have different resources.
YOU are responsible for outputs as always! If you make ToS violating content and I get aware I WILL report this.
Disclaimer
This models are shared without warranties and with the condition that it is used in a lawful and responsible way. I do not support or take responsibility for illegal, harmful, or harassing uses. By downloading or using it, you accept that you are solely responsible for how it is used.
Custom License Addendum: Distribution Restriction
Notice: Notwithstanding the base license selected for this model, the following restrictive terms apply:
No Redistribution: You are not permitted to host, mirror, or redistribute this model (checkpoint, LoRA, or Safetensors files) on any other platform, website, or service (including but not limited to Hugging Face, Tensor.art, or SeaArt) without explicit written permission from the creator.
Attribution & Source: This model is officially maintained only on Civitai or other platforms where I explicitly own the repository. To ensure users receive the correct version, updates, and safety metadata, please point users to the original URL.
Usage: All other rights regarding the use of the model for image generation remain as per the terms and the restrictions provided per model.
Description
Q4 High
FAQ
Comments (70)
EULER+SIMPLE With this configuration, when using 10STEP or more based on SHIFT5, is H5 + L5 correct?
Or is High4 + Low6 or H3+L7 more efficient? It's almost a 3:7 ratio.
I saw a graph converging closer to a 1:2 ratio.
I heard the ratio varies depending on the combination of SHIFT, sampler, and scheduler.
EULER+SIMPLE+SHIFT5
started roughly at a 1:1 ratio,
and as the total number of steps increased,
it felt like it gradually
converged toward an H1 : L2 ratio.
So I've used ratios like
H3:L4, H4:L5,
H4:L6, H4:L7,
H5:L10
. Is this appropriate?
Or should I just use a H:L = 1:1 ratio?
I'm curious.
In my experience, 2+2 actually gives the cleanest results. While the theory suggests shifting toward a 1:2 ratio as steps increase, keeping it minimal often prevents over-processing. At a low step count, this 2+2 balance maintains the best structural integrity and clarity without introducing unnecessary artifacts.
@fadedninnaย I already know that in a total of 4 steps, the ratio is 2:2.
What I want to know is the appropriate ratio at higher step counts.
When I mentioned 1:1, I was referring to the ratio, not the number of steps.
As already stated, 2 steps + 2 steps = a 1:1 ratio.
My question is about the ratio when the total number of steps exceeds 7.
In cases like SIMPLE + SHIFT8, or BETA57 + SHIFT5, the ratio generally stays close to 1:1.
However, with SIMPLE + SHIFT5, Iโve seen graphs where the High portion gradually decreases, causing the convergence value to change significantly.
Since I usually work within the 7โ12 step range, I want to know what the appropriate ratio is for that range.
Roughly speaking, if we take L5 + H10 = 15 steps as a reference, it seems that as the total step count decreases, the ratio gradually converges toward 1:1 while scaling down.
In practice, it appears to converge to something like H4 : L7 or H3 : L5.
For reference, I donโt use 2 + 2 at all.
Two steps are too short and can cause issues with motion.
Even at the lowest end, I use H3 + L4 as the minimum setting.
@dkjdjswwย I wonder if anyone would provide a appropriate answer. Because any step above 8 on distilled models introduce misbehaving or artifacts. Not to mention that in almost 99% of all cases even any step above 6 is not adding any quality to the outcome.
High steps define the motion and low refiner the details.
It is totally dependent on the noise/latent ratio when any high steps will not provide anymore to the motion. What results in almost all cases in a 1:1 ratio. In some minor cases it can benefit from more steps in low than high, but that's rare and high needs at least 2 steps in high to finish the motion.
@darksidewalkerย Thank you so much for the great quants and insight!
So almost best quality can be obtained by H3 + L6? or what do you mean exactly by not needing more than 6 steps? do you mean overall? or H6 +L6?
@omarroshdi112ย 6 total
@darksidewalkerย @omarroshdi112 I haven't analyzed it myself or reached any conclusions. Instead, I watched comparison videos on other YouTube channels and Reddit using the same light LORA. The quality needed to exceed total step 10 to show decent detail, they recommanded 15step+, and it wasn't about the DASIWA model. Personally, I tested the DASIWA v8 model with SHIFT 8 at 5:5/4:4/3:3. While the difference wasn't huge, 5:5 showed slightly better detail when zooming in quickly on freshly rendered skin right after undressing. 3:3 was the weakest. It wasn't a huge difference, but there was definitely a difference. Also, since this was SHIFT8, I used a 1:1 ratio, but I've heard SHIFT5 is best. However, I've seen many common explanations on YouTube and Reddit stating that in Euler Simple Shift5 environments, you shouldn't use 1:1 at high steps. Honestly, I didn't find any major issues with H3 L3 either... But across various communities, H2:L2 was rarely used. If it's confirmed that high steps aren't necessary, I plan to use only low steps whenever possible from now on.
@darksidewalkerย @dkjdjswwย Thank you guys so much, Really appreciate your work mate! @darksidewalker
Excellent and the best base model as always, thanks a lot !!!
Superb model with a lot of movement. But I have an issue, sometimes the camera moves and shift when I don't want to. Is there a way to make the camera static? Prompt aren't helping.
thats the biggest problem bro same in wan 2.2 i have tried every model, but that real thing is with cfg 1.0 we can't use negative prompt such as zoom in or close up
@kumarkishank959811ย You can go TrueVision without lightning or NAG
@darksidewalkerย How do you do that? Would doing 1.1 cfg be better?
My workflow has NAG feature or you use the TrueVision model from me, there you can go wuth cfg 3.5, but do not have ligthspeed.
Lightspeed/distilled models work with cfg 1, you can try 1.1 but I can not tell you if this will give quality results or introduce other problems.
@darksidewalkerย Yeah 1.1 doesn't do anything. Do you have that issue with q8 about camera drifiting and moving? Maybe it's a sampler/scheduler thing.
@SolHelย Well, every sampler+scheduler will do other results. Normally you want to try different prompting till you meet your result.
Camera is FIXED , IP-camera, Security camera, CCTV
After achieving a similar effect, I saw good results.
I'm not sure if it's completely effective, but I've seen great results.
Please specify a particular camera.
I've written everything down.
Try adding a tripod or other .
Conversely, when good movement is required, I write โDrone camera.โ
I don't know how effective it is, though.
Nobody ever does a basic video tutorial. This is so overwhelming
There are hundreds on YT for using WAN 2.2, what do you missing?
Go see Gordon Ramsay and complain that no one will teach you how to make Beef Wellington.
There are plenty of people like you in Korea too. Mostly because it's not an English-speaking country, there are absolutely many people who find it impossible to understand how to use it or even access it.
But there are so many people on YouTube teaching the exact how-to.
These two groups never see each other.
They live on different dimensions.
The fundamental gap in basic comprehension and the wildly different final world people vaguely imagine is why I think no one online honestly teaches the method.
If you imagine a world where everything gets solved with just one mobile app install button, it seems like there won't be a solution for at least the next three years. . GPUs a
In the Korean community,
when introducing the WAN 2.2 model and its results,
most people are pessimistic and criticize it..
They console each other by saying they'll find a worldview far superior to WAN.
Most have completely given up on WAN 2.2. They're currently immersed in GROK, producing illegal videos, and have become utterly despairing since censorship began.
It seems less than 1% of Koreans can actually do WAN.
Even among RTX5090/RTX5080 owners, most don't use AI. They don't even know how.
First off, the entire COMFYUI installation and usage process is in English, making it hard to understand.
For those with lower intelligence, it's like entering the gates of a new hell. Korean RTX5080/5090 users are people who just want to relax comfortably. They lack the courage to navigate this new hell.
A significant number can't even reinstall Windows themselves.
That's why most people are paying for GROK.
@dkjdjswwย nice insights, but if I would do a explanation video it would also be in English and I wrote a huge guide on WAN 22 here. So this would not cater to that people.
@dkjdjswwย ๋ ์จ๋ฐ ํ๊ตญ์ธ์ด๋ฉด์ ์ธ๊ตญ์ธ์ธ์ฒ ํ๋ฉด์ ํ๊ตญ์ธ ์ํ๊ณ ์๋ค?
@darksidewalkerย I'm comfortable with this sort of thing now, but when you're first starting off it's extremely overwhelming. Even 'basic' tutorials are usually outdated, and feel like they start in the middle instead of the beginning. Personally I benefit from still images, like screenshots of actual workflows. I think the op may be referring to something similar, which is what I had to really search for when starting out.
In WAN 2.2 i2v, when the vertical resolution exceeds 1504, a consistent issue occurs where newly appearing characters become โgiant-likeโ or develop extra limbs (e.g., four arms).
More precisely, this issue begins to appear gradually from vertical resolutions above 1440, and becomes clearly visible at 1504 and above.
This problem does not occur in WAN 2.1, and it also does not occur at standard landscape resolutions such as 1920ร1080.
The issue is specific to vertically elongated resolutions.
Although the PC hardware is fully capable of generating 1080ร1920 outputs, this bug severely limits practical usage at the moment.
When a characterโs full body is already visible from the beginning, the issue tends to appear less frequently. However, changing the characterโs pose can suddenly trigger the โgiantโ distortion.
The problem most commonly occurs when:
A new character appears, or
Body parts that were hidden in the start frame must be newly synthesized and revealed
In these cases, severe distortions occur.
However, when both Start and End frames are used, if at least one accurate intermediate forward step frame is provided, the distortion in that region is noticeably suppressed.
I have been actively using this workaround recently, particularly with QWEN Multi-Angle.
Alternatively, using only an End frame at the beginning may also work in some cases.
END FRAME ONLY for preparing Start Frame
Another workaround is:
Reusing the newly generated frame as the End frame for the next clip,
Providing it as a guideline for subsequent generation
This approach generally eliminates the distortion.
However, whenever WAN is required to fully invent and synthesize a scene on its own, the giant-like distortion reappears.
This issue is almost universally reproducible across the entire WAN 2.2 model lineup, and does not exist in WAN 2.1.
Do you know about this problem?
The resolutions you mention are not supported by WAN AI and I think you are doing something weird, because consumer grade hardware can not render that resolution without OOM.
No, I am using 1920ร1080 and 1080ร1920, or resolutions very close to them, without any problems at all.
https://www.youtube.com/watch?v=wJAR1j_wTns
https://www.youtube.com/watch?v=mfNlsAqVd3o
THis Youtube channel is not my main channel, I made this channel just recently for upload some video clip only for NSFW test on YUTUBE.
These resolutions themselves are not an issue;
this is a bug that occurs only under specific conditions, and it does not exist in version 2.1.
Please check the video.
This was generated in WAN at 1920ร1080 and 1080ร1920 without any upscaling.
After that, it was upscaled to 4K. (from native 1080p video by wan 2.2)
I am not a beginner.
I have a very high level of experience and have generated over 200,000 WAN videos.
summary
Wan 2.1 = 1080x1920 and 1920x1080 noproblem
wan 2.1 = 1920x1080 / 1600x1280 /1600x1312 =100% no problem this range I allways use this area resolution. 1.8MP~2.08MPixtel. native resolution.
it doesn't need expensive GPU
of course 81frame+ and Q8 gguf Dasiwa model.
I am using 81~97fps for 1.8MP~2.1MP RANGE (FHD=2.08MP)
Torch2Multigpu nodes and 24GB GPU +64GB RAM is enough to make native 1920x1080p video
this node is not compatible with comfyui v8.xx maybe.
so I am using COMFY v7.00 now.
this node is better than block swap. you can produce 1920x1080 video with only 3090 GPU
6step = 30min
5step = 25min
4step =20min
for 81 frame. 100% clear and good. 85fps~97fps range is also possible.
and RTX 5090 is about 3.5 times faster than 3090.
It's almost factory-level performance.
I have few RTX 5090 and single 3090.
3090 also can produce 1920x1080 81frame+ video.
1440x1440 = small bug problem , but I con control (1440 height has a small problem)
1120x1680 = good for 2:3 ratio Grok NSFW picture. but Need some control too
you can see vertical video in my channel , the good result for long video . minimum movement and end frame.
you can check it the youtube channel,
Minimized movement and controlled a bug-like issue that only appears in vertically long videos using the end frame. The video is quite long... I natively stitched together about 800 frames. Afterward, I doubled the playback time to match the music and upscaled it from 1080p to 4K.
I expected it, but it's surprising how many people haven't tried 1920x1080.
With 24GB VRAM and specific nodes, it's done in no time.
You'll need at least 64GB of RAM, and it might even require 80GB.
Now RAM is just too expensive.
I can't say it's easy anymore.
But until a few months ago, it was incredibly easy and cheap.
Wan 2.2 1080p 6step Simple Setup
https://www.youtube.com/watch?v=frLJCkDyvLE
Because WAN only supports multiples of 16, you have to choose between 1072p and 1088p, and usually 1072p with letterboxing is used.
Alternatively, you create something closer to a 16:10 or 2:3 aspect ratio (for example, 1680ร1200), and then, depending on the situation, expand it further using VAE outpainting.
I canโt say for sure, but for 5-second generation, it feels like 24 GB of VRAM is practically a required condition.
With 24 GB VRAM, it already gets almost completely filled, and system RAM usage is also very high.
I havenโt tried block swap yet.
On 24 GB, it seems possible up to about 89 frames for now.
Usually, I just reduce the pixel count slightly and go for 97 frames or more.
Landscape videos in square-ish aspect ratios are extremely easy to produce and have no issues at all.
Only vertical videos have many bug-like, difficult issues in WAN 2.2.
But I realized that there are almost no creators working in the exact same niche as I am.
I should probably try again with a model that incorporates elements from the recently released 2.1.
For me, producing 1080p videos with WAN has been a completely solved, finished problem for a long time.
The only issue is the bug-like problems with vertical aspect ratios in 2.2.
Even those are gradually being figured out in terms of the conditions to่งฃ้ค them, and if you look at the video above, itโs already rendering with those issues resolved.
Even 64 GB of RAM doesnโt feel particularly sufficient.
Especially with this setup, depending on the settings, there are cases where the TEXT encoder fails to properly switch to FP16.
On the other hand, the UNet processing stage actually has more headroom and runs more smoothly.
I also found a model that fixes several issues that appear only in 2.2.
For example, problems like stuttering during motion.
They said it partially inserts blocks from 2.1.
Because of that, I was hoping for 2.5, but it looks like Iโll have to wait for 3.0 instead.
And if 3.0 is eventually released, Iโll probably have to buy a completely new, next-generation GPU as well.
not 16x, 8x is OK, I can output 360x640 here
a little late to comment with how much i've been using it already, but this checkpoint has been fantastic so far coupled with the FastFidelity C-AiO workflow. you put out excellent easy to use resources for people like me who want to create aesthetically refined pieces but don't have the time to make generation their primary focus. thank you!
Damn this is looking amazing I have to try it XD
Dumb question to be asked, but is mixing a Q2 High with Q4 Low viable strategy?
As i understand High controls motion, Low controls detailing, my system bearly can run Q4H+Q4L so i'm simply wondering if connecting Q3H or Q2H with Q4L is possible idea.
If you ask, "why don't you test it yourself" well guys, test require lot of samples, meanwhile it takes 2 hours to render 9 seconds in 12fps for my set up.
So here i am, asking, if there been someone experimenting on this, while having better setup than 4+32gb.
And one more question, how much better is V9 compared to V8? Is it more precise, or rather more efficient? Where and in what is the upgrade? (since it is impossible for me to experiment to experience the difference)
Why I even bother - you may ask, I simply wish to experience new technology, and "try my best" or learn new techniques.
For ex. instead of using one prompt generator i learned and analyzed and made multi prompt smooth flow video generator ยฏ\_(ใ)_/ยฏ pointless exercise given how time consuming it is and given that 3 scene (or 3 step, call as you wish) generation nearly resulted in memory overflow and crash. (acc happen on 1st try, had to close everything to free as much ram possible for 2nd try)
I would like here to thank the Author, for making those smaller Q models so some fool, that did not upgraded PC when was time for it, can acc toy around with modern tools :) o7
Limited time, only one test group: High at Q2, with Low variants at Q2, Q4, and Q8.
@PotatoChipOwOย Thank you :) That helps
Perfect, used your TastySin model before and loved it, this one is even better <3
This is my favorite WAN checkpoint so far. Excellent results.
Thank you!
ๆๆฒๆณๅฐๆไฝฟ็จQ8้ๅใๅจๆๅจๆ็4060๏ผ8GB้กฏๅญ๏ผๅฑ
็ถ2+2ๆญฅ๏ผ512*720๏ผ81ๅน๏ผๆกๆจฃๅช่ฑ180็ง! !
้็ถ่ชชๆ็่จๆถ้ซ32GBๅ
จ่ขซๅๆปฟใไฝ้ๆจกๅๅฏฆๅจๅช็ง๐ฅฐ๐ฅฐ
็็, ้ๅบฆๅพๅฟซ.
Hi, where can i download fp8 version of this model ? i searched around but can't find it
You searched? Here are some ways to find it:
- suggested resources just below
- my collection, right panel
- my page, promoted model's
- civitai search, the name of the model
...
But here is the direct link https://civitai.com/models/1981116
@darksidewalkerย Thank you very much, I dont know how i missed it. Thank you for helping me.
Thanks a lot for your workflows and checkpoints
If you are starting like me, it's all you really need at the moment
Keep up the good work !
lightx2v_I2V_14B_480p_cfg_step_distill_rank256_bf16 Can I use this?
I do not know if you can, but it is possible. Is it also advised? - No.
Distillation is already baked in.
This user's models are among the best; so far, they've given me the best results with my Asus RTX 5070 Prime. Besides, as my friend mentioned, the workflows are also optimized. The video quality and execution time are excellent; 5-second videos take a maximum of 4 minutes. Therefore, FP8 FastSpeed โโand Guff run relatively quickly for me, while those with higher step counts, like 30 and 40, are really slow and cumbersome. Excellent work.
Hello, how do I choose between SynthSeduction and TastySin? maybe Q4
I draw realistic figures with background scenery.
SynthSeduction should be better for that.
This model is brilliant! The best so far!! I appreciate you and your skills my friend!!!
Very kind, thank you!
Thats awesome! Great job
where is the q3 high model ?
In front of you right under the headline.
damn for some reason my adblock was HIDING only that one :) And here I thought he forgot to upload q3 high for some reason :)
Crazy XD
I had been fighting with another mix for weeks, trying to stop the people from dancing all the time, or thrusting into each other without prompting.
Used this yesterday and it was surreal, they moved when I told them to move and they did what I told them to do. No dancing, no thrusting, it was fantastic.
I wish I had discovered your models sooner, thank you for sharing.
Thank you! Kind words :) I hope you enjoy it!
@darksidewalkerย I have used this model for a few days now and the actions are still doing pretty much what I ask them. The problem though, and it's turning out to be a big problem, is the infamous Wan2.2 slow motion. What it means is that if you create 30 seconds of video, you're only getting 15 to 20 seconds of animation because everything is running slowly. If I switch back in the previous mix, everything is fast again but I get the annoying movements. Have you found any reliable solutions to this problem at all?
@youtougle2422ย WAN 2.2 does not create 30s of video, it will be a mess, it is trained on 5s.
The problem with slow motions in some situations comes from distillation (4 step). If you totally want to prevent this you will have to use a model like my TrueVision or basic WAN 2.2, without distillation and 20-30 steps.
@darksidewalkerย Sorry I should have mentioned, I am using an SVI looper workflow so I am chaining several smaller clips together, not just trying to generate a single 30 second clip. I notice that other loras can affect things as well. I will see if I can integrate the Painter node into the workflow which I use to keep the speed up in my standard Wan2.2 workflow.
Please don't take my comment as being something negative though, I still prefer what it creates and the motion reliability is far more important than the slowdown. Wan2.2 has just been frustrating with the slowdown problem, especially for a noob to AI gen like myself. I genuinely appreciate the time you spend creating and sharing these models.
thank you very much
q6 high gguf seems broken, comyui says cannot reshape array of size 12547712 into shape (5120,4200)
q6 low works fine.
any chance of nvfp4 or fp4?
Planned for next version. Be aware that this would be less quality than Q5.
@darksidewalkerย it's supposed to be like Q4 but 'better' i guess, thanks! i will wait for it :D
today i'm do stupid update "comfyui+python_dependencies" and now my work is broke. (it's run flash process but nothing even got this text in comfyui console. please help.
-------------------------------
got prompt
Failed to validate prompt for output 1512:1588:316:
* GetImageSize 1512:1588:328:
- Required input is missing: image
Output will be ignored
Failed to validate prompt for output 9:
* VHS_VideoCombine 28:
- Required input is missing: images
Output will be ignored
Failed to validate prompt for output 28:
* (prompt):
- Required input is missing: images
Output will be ignored
Failed to validate prompt for output 1512:1588:318:
Output will be ignored
Prompt executed in 2.02 seconds
--------------------------------
Change the codec from VHS node to one that works for you (e.g. h264) or install ffmpeg properly.
@darksidewalkerย changed or ffmpeg reinstall, same error result. anyway Thank you for guide me.
It keeps turning underwear to be crotch-less (revealing genitals)... Anyway to avoid this? I've tried prompting their genitals to remain covered, but unsuccessful.... ok i kinda solved it by completely avoiding the term "crotch"
Bradar this is gold!! My RTX 2080Ti with 11gb vram can finally generate videos thank you bradar
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.